Try to fix pg_upgrade test on Windows, again.

Further to commit 54e72b6, if rmtree() fails while cleaning up in
pg_upgrade, try again.  This gives our Windows unlink() wrapper a chance
to reach its wait-for-the-other-process-to-go-away logic, if the first
go around initiated the unlink of a file that a concurrently exiting
program still has open.

Discussion: https://postgr.es/m/CA%2BhUKGKCVy2%3Do%3Dd8c2Va6a_3Rpf_KkhUitkWCZ3hzuO2VwLMXA%40mail.gmail.com

Author: macdice

src/bin/pg_upgrade/util.c

@@ -68,7 +68,12 @@ cleanup_output_dirs(void)
 	if (log_opts.retain)
 		return;
 
-	(void) rmtree(log_opts.basedir, true);
+	/*
+	 * Try twice.  The second time might wait for files to finish being
+	 * unlinked, on Windows.
+	 */
+	if (!rmtree(log_opts.basedir, true))
+		rmtree(log_opts.basedir, true);
 
 	/* Remove pg_upgrade_output.d only if empty */
 	switch (pg_check_dir(log_opts.rootdir))
@@ -80,7 +85,13 @@ cleanup_output_dirs(void)
 
 		case 1:					/* exists and empty */
 		case 2:					/* exists and contains only dot files */
-			(void) rmtree(log_opts.rootdir, true);
+
+			/*
+			 * Try twice.  The second time might wait for files to finish
+			 * being unlinked, on Windows.
+			 */
+			if (!rmtree(log_opts.rootdir, true))
+				rmtree(log_opts.rootdir, true);
 			break;
 
 		case 4:					/* exists */